Voyage AI

Converts unstructured text into dense vector representations using the voyage-3.5 model, supporting up to 32K tokens of context per input. The model is optimized for retrieval-augmented generation (RAG) pipelines and produces 3x-8x shorter vectors than competing embeddings while maintaining superior accuracy on benchmark tasks. Handles arbitrary text length by chunking internally and returning normalized vector outputs compatible with any vector database.

Provides the voyage-3.5-lite variant, a compressed version of the general-purpose embedding model optimized for inference speed and reduced computational requirements. Maintains competitive accuracy on retrieval benchmarks while consuming 4x less compute resources, enabling deployment on edge devices, serverless functions, and cost-constrained environments. Produces the same vector format as voyage-3.5 for seamless integration into existing RAG pipelines.

Voyage AI embeddings and reranking models are designed to integrate with any large language model (OpenAI, Anthropic, Ollama, open-source LLMs, etc.) without vendor-specific adapters. The embedding and reranking outputs conform to standard formats that any LLM can consume, enabling flexible RAG pipeline composition. Organizations can combine Voyage embeddings with their choice of LLM without architectural constraints or proprietary integrations.

Provides specialized embedding models fine-tuned for specific domains (finance, legal, code) that outperform general-purpose embeddings on domain-specific retrieval benchmarks. Each model is trained on domain-relevant corpora and optimized for terminology, context, and semantic relationships unique to that field. Integrates seamlessly into RAG pipelines by replacing the general-purpose embedding model while maintaining the same vector database interface.

Enables organizations to request custom fine-tuned embedding models tailored to their proprietary data, terminology, and domain-specific requirements. The fine-tuning process leverages Voyage AI's base models and adapts them to company-specific semantic relationships, enabling superior retrieval performance on internal knowledge bases and proprietary corpora. Custom models are deployed via the same API interface as standard models, requiring no changes to downstream RAG infrastructure.

The voyage-multimodal-3.5 model generates embeddings for both text and images in a shared vector space, enabling cross-modal retrieval where text queries can retrieve relevant images and vice versa. The model is trained to align text and image semantics, producing vectors that preserve both modalities' semantic relationships. Integrates into RAG pipelines to support hybrid document collections containing both text and visual content.

The voyage-context-3 model generates embeddings that preserve both chunk-level details and global document context, addressing the limitation of standard embeddings that lose document-level semantics when chunking. The model is trained to understand how individual chunks relate to the overall document structure and meaning, improving retrieval accuracy for systems that chunk documents into smaller units. Outputs embeddings compatible with standard vector databases while maintaining awareness of document-level context.

The rerank-2.5 model re-orders retrieved search results to improve relevance ranking, using instruction-following capabilities to adapt reranking behavior based on user intent. The model takes a query and a list of candidate documents, scores each document's relevance to the query, and returns a ranked list optimized for precision. Integrates into RAG pipelines as a post-retrieval step to refine results from vector database queries before passing to the LLM.

The rerank-2.5-lite variant provides a compressed reranking model optimized for inference speed and reduced computational requirements, enabling real-time reranking in latency-sensitive applications. Maintains competitive ranking accuracy compared to rerank-2.5 while consuming significantly less compute resources, making it suitable for high-throughput retrieval pipelines and edge deployments. Produces the same ranking output format as rerank-2.5 for seamless pipeline integration.

Provides a batch processing API for embedding and reranking large volumes of documents asynchronously, optimizing for throughput and cost efficiency over latency. The batch API accepts bulk requests, processes them in optimized batches, and returns results via callback or polling mechanism. Enables cost-effective processing of millions of documents without hitting rate limits or incurring per-request overhead of synchronous API calls.

Voyage AI embeddings are designed to be compatible with any vector database (Pinecone, Weaviate, Milvus, Qdrant, etc.) without custom adapters or format conversions. The API returns standard dense vectors in normalized format that conform to vector database input specifications, enabling plug-and-play integration. Organizations can switch between Voyage embedding models or migrate to other providers without modifying vector database schemas or retrieval code.